Real-time seismic response prediction for electrical equipment using CNN-MLP model

Abstract. Monitoring the seismic response of substation equipment is critical for developing effective real-time emergency response strategies. However, traditional monitoring approaches face significant challenges due to the electromagnetic sensitivity of electrical equipment. This study proposes a CNN-MLP neural network architecture that aggregates both long-duration seismic data and local temporal window information to enable real-time prediction of equipment acceleration responses based on seismic accelerations. Additionally, a beyond-range training strategy is introduced to enhance model performance under high-intensity seismic conditions. The proposed model is evaluated using a damped ±800kV wall bushing, which has been validated through shaking table tests, and compared to an LSTM model. Results show that the CNN-MLP model significantly outperforms the LSTM model, with the beyond-range training strategy effectively improving prediction accuracy and stability. This method requires only ground motion signals as input, avoiding the need to install sensors on electromagnetically sensitive equipment surfaces, thereby providing reliable support for emergency decision-making in substations during earthquakes.